Using Logistic Regression and a Novel Machine Learning Technique to Predict Discharge Status after Craniotomy for Meningioma

Introduction: Discharge disposition is a significant consideration for patients and providers deciding whether to undergo surgical treatment. Disposition impacts decisions about timing of surgery, time away from employment for patients and their caregivers, and choice of a surgeon or hospital. This information is also important to providers, who need to navigate the logistics of hospital bed availability and work with insurance companies to ensure appropriate coverage for surgical candidates. Despite its importance, no attempts have been made to predict discharge status following meningioma resection.

Traditionally, logistic regression analysis has been used to model medical outcomes. However, focusing on logistic regression ignores other classes of machine learning algorithms, which have the potential to yield novel clinical insights. We take advantage of these lesser-known algorithms to make predictions in an entirely unique way: we train and rank a range of different algorithms, selecting the most predictive algorithms for inclusion in an ensemble model, which combines several machine learning algorithms to take advantage of the individual strengths of the different model classes. We also train algorithms using hold out datasets and cross-validation, allowing for the inclusion of a large number of predictors without overfitting. Here, we use both logistic regression and our machine learning approach to predict discharge disposition after meningioma resection.

Methods: We completed a single-center, retrospective study of 552 patients undergoing craniotomy for meningioma from 2001 to 2015. Multivariable logistic regression was used to predict disposition status at the time of hospital discharge. A predictive model was constructed and underwent calibration and discrimination.

34 machine learning algorithms were trained to predict discharge disposition and the three most predictive models were combined with an Elastic Net to create an ensemble model. Permutation and partial dependence analysis was performed to identify how important variables impact the model.

Results: Presence of tumor edema, tumor size, presence of a preoperative motor deficit, and age were significant and included in the logistic regression model. Model calibration and discrimination were assessed using the Hosmer-Lemeshow test (p = 0.05) and Receiver Operating Characteristic (ROC) curve (area under the curve [AUC] = 0.77; bootstrapped 95% CI 0.73-.82), respectively.

84 predictors were used to create an ensemble model composed of an Elastic-Net Classifier, Nystroem Kernel SVM Classifier, and Random Forest Classifier. ROC analysis showed AUC = 0.78 (95% CI 0.77-.8). Permutation and partial dependence analysis demonstrated that tumor size, presentation at the ED, convexity tumor location, and preoperative motor deficit most strongly influenced the model’s predictions.

Conclusion: Both logistic regression and an ensemble machine learning model predict discharge disposition following craniotomy for meningioma with fidelity and identify risk factors for non-home discharge. Each technique has its advantages – logistic regression is well known and well-established, and is relatively computationally simple. Alternative machine learning algorithms tend to be complex and require more expertise, but allow for novel approaches to and interpretation of data. Taken together, these are powerful techniques for modeling a multitude of neurosurgical outcomes.